{"id":"92190e85-d4a7-4bb0-a6a2-f8d03f03de01","arxiv_id":"1908.02460","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"ENFNet, an edge-guided non-local FCN, reports state-of-the-art F-measure and MAE on five salient object detection benchmarks.","lead":"This paper proposes ENFNet, a deep network that uses edge maps to guide feature learning for salient object detection, aiming to produce sharper object boundaries. The method reports improved F-measure and MAE scores over prior methods on five standard benchmarks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ablation does not isolate edge-prior benefit: adding the EGB adds capacity, and the cited edge extractor [22] is an object-proposal method, so the claimed source of gain is unsupported.","rationale":"I agree with the reader's conditional verdict but for a somewhat different reason. The reader's weakest assumption is that Edge Boxes edge maps are accurate enough across datasets. My concern is more fundamental: even if the edge maps are perfectly accurate, the experimental design cannot attribute the reported gains to the edge prior, because the 'Five EGB' and 'Three EGB' variants add substantial model capacity relative to the 'Zero EGB' NLDF baseline. A standard capacity-control ablation (e.g., adding the same number of convolutional layers without any edge-condition input) is missing. This matters because the paper's novelty and the abstract's claim of 'edge-guided feature learning' rest on the edge-specific mechanism, not merely on having a deeper network. Additionally, the reference for edge extraction is mismatched: [22] is Edge Boxes, a proposal generator, not a dense edge detector. The paper does not describe how its outputs become the 176x176 edge maps used by the condition network, making the exact input to the proposed block unreproducible. These issues weaken the support for the central mechanistic claim but do not necessarily falsify the raw performance numbers, which could still be correct. Therefore I would keep the verdict conditional, but with a sharper condition: the authors should either run the capacity/edge-prior control or release code and edge-map generation details. This does not move the overall verdict from CONDITIONAL, so I mark it UNCHANGED.","tokens_in":13685,"tokens_out":5026,"duration_ms":53859,"concrete_test":"Train the same ENFNet architecture under three conditions with identical hyperparameters: (a) real Edge Boxes-derived edge maps, (b) a constant-zero tensor of the same shape, (c) Gaussian noise, as the condition input to the guidance block; then compare max F on DUT-OMRON and SOD. If (b)/(c) match the reported 0.779/0.831 within noise, the gain over NLDF is capacity, not edge guidance. Also ask the authors to release the edge-map generation code, since [22] is an object-proposal method and the conversion to XE_i is unspecified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that edge-prior embedding via the guidance block is what lifts ENFNet over NLDF, with the largest gains on DUT-OMRON and SOD. The support for that attribution is Table III, which compares 'Five EGB', 'Three EGB', and 'Zero EGB'. But 'Zero EGB' is the NLDF baseline with none of the added condition network or affine-transform layers; the EGB variants add dozens of convolutional layers and parameters (Table I: each block has two 3x3 or 5x5 convs, plus a four-layer condition network). Adding capacity alone, without any edge input, could produce part or all of the reported +2.6%/+2.1% F-measure gains, so the ablation does not establish that edge information is the operative factor. The ambiguity is compounded by the specification of the edge prior: Section III-C says edge maps come from 'the existing method [22]', and reference [22] is Zitnick and Dollár's Edge Boxes, an object-proposal method rather than a dense edge detector; the paper nowhere states how proposal/edge outputs are converted to the 176x176 maps XE_i used by the condition network. This is a reproducibility gap at the exact point where the novelty lives. No code or error bars are provided, so the central mechanism cannot currently be verified from the manuscript alone.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ENFNet, a VGG-16-based fully convolutional network for salient object detection. The architecture extracts hierarchical multi-scale features from five side outputs, computes local contrast features by subtracting local averages, incorporates a global context branch, and uses a proposed edge guidance block (EGB) that applies a spatial affine transform conditioned on edge maps to the multi-scale features. The model is trained end-to-end on MSRA-B with a cross-entropy loss plus an IoU boundary loss, and is evaluated on five public benchmarks. The authors report that ENFNet achieves the best F-measure and MAE among 13 compared methods on all datasets, with the largest gains over NLDF on DUT-OMRON and SOD, and they include an ablation varying the number of edge guidance blocks.","tokens_in":13986,"tokens_out":4054,"duration_ms":44805,"significance":"If the reported results are robust, the paper would provide a simple and reasonably efficient mechanism for incorporating boundary priors into a saliency network, with consistent though modest gains across five benchmarks. The evaluation is conventional and the comparison set is broad, and the runtime of 0.08 s per image is a practical strength. However, the central attribution of the improvement to edge information is not established by the current experiments: the ablation confounds the edge prior with added model capacity, the edge-map generation step is both mis-cited and under-specified, and the boundary-loss formulation is not fully reproducible. No code or error bars are provided. The significance is therefore conditional on the authors closing these reproducibility and attribution gaps.","major_comments":[{"comment":"The ablation study does not isolate the benefit of the edge prior. The 'Zero EGB' condition removes the entire edge guidance block and its condition network, while 'Three EGB' and 'Five EGB' add many convolutional layers (Table I shows each block contains two or more convolutions plus the shared condition network). The comparison therefore confounds the presence of edge information with a substantial increase in model capacity. A control experiment with the same added layers but with constant or random edge input, or an ablation that removes only the edge conditioning while keeping the affine transform capacity, is needed to support the claim that edge information rather than capacity drives the reported gains.","section":"Section IV-C, Table III"},{"comment":"The paper states that edge maps are extracted using 'the existing method [22]', but reference [22] is Zitnick and Dollár's Edge Boxes, an object-proposal method, not a dense edge detector. The manuscript does not describe how Edge Boxes outputs are converted into the 176×176 edge maps XE_i that feed the condition network. This is a reproducibility gap at the exact point where the novelty of the method lies; the authors need to specify the edge extraction procedure, the preprocessing, and whether separate edge maps are computed for each of the five scales.","section":"Section III-C"},{"comment":"The IoU boundary loss is underspecified. It is unclear how the Sobel operator followed by tanh is applied to form C_j and Ĉ_j from the ground-truth and predicted saliency maps, whether these boundary maps are computed at each scale j or only at the output resolution, and how the intersection and union counts are made differentiable for backpropagation. In addition, λ_j and γ_j are introduced as per-term weights in Eq. (5) but in Section III-E they are set to the scalar value 1, leaving the multi-scale or multi-loss structure of the objective ambiguous.","section":"Section III-D, Eqs. (5)-(8)"},{"comment":"The central claim that ENFNet 'achieves the best performance on all datasets' rests on margins that are often small relative to the second-best method (e.g., 0.915 vs 0.902 max F on HKU-IS and 0.843 vs 0.831 on PASCAL-S), but the paper reports no variance, number of runs, or significance tests. Since the network is trained with a fixed seed and no code is provided, the reader cannot determine whether these differences are meaningful. The authors should report standard deviations or at least specify the evaluation protocol and training runs used to obtain the reported numbers.","section":"Section IV-B, Table II"}],"minor_comments":[{"comment":"The F-measure formula as printed has a typo: the denominator should be β²·Precision + Recall, not β²·Precision·Recall; as written, the expression is independent of Precision and Recall.","section":"Section IV-A, Eq. (9)"},{"comment":"The future-work paragraph contains unresolved placeholders '[ ?]' for semantic priors and thermal infrared data; these should be replaced with appropriate citations or removed.","section":"Section V"},{"comment":"The sentence 'we only calcute the evaluation metrics on the test dataset on HKU-IS' is unclear; the authors should specify how the HKU-IS train/test split is determined and state whether all images are used for the other datasets.","section":"Section IV-A"},{"comment":"There are typographical and specification issues in Table I: 'kernal' should be 'kernel', and the table should report the input and output tensor shapes of the condition network and how the shared intermediate condition is distributed to the five edge guidance blocks.","section":"Table I"},{"comment":"The text says the local feature XL has size 176×176 and is then upsampled to 352×352 for fusion with the global feature, while the network output is stated to be 176×176; the final resolution and the role of the 352×352 feature should be clarified.","section":"Section III-B"}],"recommendation":"major_revision","confidential_remarks":"This is a standard empirical architecture paper that fits the journal's scope. My main concern is not the plausibility of the results but the support for the central attribution claim: the ablation does not control for capacity, and the edge prior generation is both mis-cited and underspecified. I would encourage the editor to require either code release or a revised ablation that holds capacity fixed while toggling the edge information, together with a precise description of the edge map computation and the IoU boundary loss. With those revisions, the paper could become acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my take on arXiv:1908.02460. The short version: ENFNet is an incremental but sensible extension of NLDF, borrowing the spatial feature transform idea from image super-resolution to inject edge priors into saliency features. The idea is clean and the experiments are standard; the numbers genuinely improve on NLDF across five benchmarks (e.g., +2.6% F-measure on DUT-OMRON, +2.1% on SOD). That part looks legitimate.\n\nBut the central attribution is shaky. The ablation in Table III compares Five EGB against Zero EGB, and Zero EGB is just NLDF with none of the new modules. The edge guidance blocks add a bunch of convolutional layers and a condition network, so the gain could come from added capacity alone. There's no control with a non-informative edge map or with the condition network detached from any edge input. Without that, you cannot conclude edge information is what helps. This is the load-bearing claim of the paper, and the evidence doesn't isolate it.\n\nAlso, the edge prior itself is underspecified. The paper cites [22] as 'EdgeNet,' but [22] is Zitnick and Dollár's Edge Boxes, an object-proposal method. Edge Boxes uses structured edge detection internally, but the paper never says how the dense edge map is produced, normalized, or resized to the 176x176 input XE_i. That's a reproducibility gap at exactly the novel component.\n\nSmaller issues: no code, no error bars, and Eq. (9) for F-measure is a typo (the denominator should be β² P + R, not β² P R). The boundary loss via Sobel and tanh is described in prose only; a formula would help. The comparison set is from an earlier generation, but for a method paper of this vintage, that's forgivable.\n\nWhat the paper does well: it adapts SFT to a new task sensibly, applies it hierarchically, and produces consistent, credible improvements. The idea is not a big leap, but it's a real one. The writing is mostly clear.\n\nI'd send this to peer review rather than desk reject. The method deserves referees' time, but the authors need to produce a proper ablation isolating the edge prior and specify their edge extraction pipeline. If I were the editor, I'd ask for major revision on those two points. The core method probably works; the evidence for why it works is what's missing.","headline":"A plausible incremental extension of NLDF with a clean edge-guidance idea, but the ablation does not isolate the edge prior from added capacity.","tokens_in":14481,"tokens_out":4228,"would_cite":false,"duration_ms":42443,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"ENFNet claims that embedding edge priors into hierarchical FCN features via a spatial affine transform yields state-of-the-art salient object detection on five benchmarks.","keywords":["salient object detection","edge guidance","non-local features","fully convolutional network","boundary preservation","feature-wise affine transform","EdgeBoxes","benchmark evaluation"],"falsifier":"Compare ENFNet against its zero-edge-guidance variant on a subset of images where EdgeBoxes returns no edges or edges that disagree with ground-truth object boundaries: if the edge-guided model does not beat the no-guidance variant there, the reported gains come from other components. A direct experiment could replace the EdgeBoxes maps with random or inverted edge maps and check whether the F-measure improvements persist.","tokens_in":13532,"feed_emoji":"🖼️","tokens_out":5831,"duration_ms":58110,"temperature":0.7,"pith_summary":"The paper tries to establish that feeding explicit edge-map priors into a fully convolutional network at every feature scale produces salient-object maps with sharper boundaries, and that this design beats thirteen existing methods on five standard benchmarks. The proposed network, ENFNet, takes a non-local FCN baseline, computes edge maps with Edge Boxes, and uses a guidance block that applies a per-pixel affine transform to each hierarchical feature map. Trained on one dataset (MSRA-B) and tested on five others, the method reports the best F-measure and MAE scores everywhere, with its largest F-measure gains over the runner-up being +2.6% on DUT-OMRON and +2.1% on SOD. If the claim holds, it means boundary preservation in saliency detectors can be improved by a simple, modular edge-guidance mechanism rather than a redesigned backbone.","feed_headline":"Edge-guided network sharpens salient object boundaries","feed_subtitle":"Edge priors embedded at every scale help an FCN beat 13 prior methods on five saliency benchmarks.","key_machinery":"The central machinery is the edge guidance block, a spatially feature-wise affine transform. A condition network built from four $3 \\times 3$ convolutional layers turns an EdgeBoxes edge map into per-scale features, from which two branches predict scale maps $\\gamma$ and $\\beta$. The transform $\\mathrm{EGB}(X_i \\mid \\gamma, \\beta) = X_i \\odot \\gamma + \\beta$ embeds the edge prior into each hierarchical feature map. The same block feeds contrast features $X^c_i = X^F_i - \\mathrm{AvgPool}(X^F_i)$, which capture local foreground/background differences, and the deconvolution path that fuses local and global features. The IoU boundary loss ties the edge guidance to sharp predicted boundaries.","core_discovery":"The paper claims that embedding edge prior knowledge into hierarchical FCN features fixes the blurred-boundary weakness of saliency detectors. The proposed ENFNet builds on the non-local deep FCN of [20], computes edge maps with EdgeBoxes [22], and passes them through a condition network that predicts per-pixel affine parameters ($\\gamma$, $\\beta$). Each side-output feature $X_i$ is transformed by $\\mathrm{EGB}(X_i \\mid \\gamma, \\beta) = X_i \\odot \\gamma + \\beta$, so the edge guidance acts both feature-wise and spatially. Downstream, local contrast features are formed by subtracting each edge-aware feature from its local average, global context comes from the top of the network, and a cross-entropy plus IoU boundary loss trains the whole model. Trained only on MSRA-B and evaluated on HKU-IS, PASCAL-S, DUT-OMRON, ECSSD, and SOD, the method reports the best F-measure and MAE among thirteen compared methods, with the largest F-measure gains over the second-best method being +2.6% on DUT-OMRON and +2.1% on SOD.","pith_inferences":["The paper leaves implicit that the same affine edge-guidance block is a generic conditioning mechanism: it could be plugged into other dense prediction networks (semantic segmentation, instance boundaries) wherever an external boundary prior is available.","Because the edge maps come from an off-the-shelf proposal method rather than a learned edge detector, an end-to-end version that learns edges jointly might close the gap further; this is an extension, not something the paper tests.","The reported gains (under 3 F-measure points) suggest the main effect is boundary refinement rather than discovery of new salient regions; a test that isolates boundary-pixel accuracy, such as F-measure on only boundary bands, would show whether that is the real source.","The method's dependence on EdgeBoxes means its generalization on image types where EdgeBoxes produces few edges, such as low-contrast natural scenes, is an open question rather than established behavior."],"forward_implications":["Saliency detectors built on FCN backbones can recover boundary sharpness without replacing the backbone, by inserting edge-guidance blocks at multiple scales.","The guidance block is cheap enough that the full model retains a 0.08-second per-image runtime on a single consumer GPU.","Using edge guidance in all five hierarchical scales beats using only three scales or none, so the hierarchical embedding itself contributes to the reported gains.","The improvements concentrate on difficult datasets (DUT-OMRON, SOD), where cluttered backgrounds and low contrast make boundaries hard, so edge guidance matters most where boundary cues are scarce.","The reported MAE reductions, such as 15% and 9% relative decreases over the runner-up on DUT-OMRON and SOD, imply the method not only localizes objects but also predicts saliency values closer to ground truth."],"supporting_citations":[{"why":"Supplies the non-local FCN baseline whose architecture ENFNet extends and the primary numbers it must beat.","marker":"[20]"},{"why":"Produces the edge maps that serve as the boundary prior fed into the guidance block.","marker":"[22]"},{"why":"Introduces the deep spatial feature transform (affine conditioning) that the edge guidance block adapts from segmentation priors to edge priors.","marker":"[23]"},{"why":"Provides the edge-aware comparison method Amulet, an important reference for evaluating edge-guided design.","marker":"[18]"},{"why":"Provides the MSRA-B training set used to train ENFNet, so the benchmark comparisons rest on this training data.","marker":"[52]"}],"fun_headline_variants":["Edge priors sharpen saliency boundaries in FCNs","Edge-guided FCN surpasses prior art on five datasets","Non-local features plus edge guidance refine salient object detection","ENFNet edge-guided non-local network for sharper saliency"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim assumes that the edge maps produced by Edge Boxes are accurate enough, across all test images, to serve as trustworthy boundary priors; on low-contrast or cluttered images where the edge detector fails, the guidance could mislead the network.","fun_headline_variants_meta":{"raw":{"variants":["Edge priors sharpen saliency boundaries in FCNs","Edge-guided FCN surpasses prior art on five datasets","Non-local features plus edge guidance refine salient object detection","ENFNet edge-guided non-local network for sharper saliency"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000713,"raw_usage":{"total_tokens":3209,"prompt_tokens":951,"completion_tokens":2258,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":2191}},"tokens_in":567,"tokens_out":2258,"duration_ms":16988,"temperature":1.0,"reasoning_tokens":2191,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:42:35.470645+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare ENFNet against its zero-edge-guidance variant on a subset of images where EdgeBoxes returns no edges or edges that disagree with ground-truth object boundaries: if the edge-guided model does not beat the no-guidance variant there, the reported gains come from other components. A direct experiment could replace the EdgeBoxes maps with random or inverted edge maps and check whether the F-measure improvements persist.","supporting_citations":[{"cited_title":"Non-local deep features for salient object detection,","cited_arxiv_id":null,"evidence_quote":"Supplies the non-local FCN baseline whose architecture ENFNet extends and the primary numbers it must beat."},{"cited_title":"Edge boxes: Locating object proposals from edges,","cited_arxiv_id":null,"evidence_quote":"Produces the edge maps that serve as the boundary prior fed into the guidance block."},{"cited_title":"Recovering realistic texture in image super-resolution by deep spatial feature transform,","cited_arxiv_id":null,"evidence_quote":"Introduces the deep spatial feature transform (affine conditioning) that the edge guidance block adapts from segmentation priors to edge priors."},{"cited_title":"Amulet: Aggre- gating multi-level convolutional features for salient object detection,","cited_arxiv_id":null,"evidence_quote":"Provides the edge-aware comparison method Amulet, an important reference for evaluating edge-guided design."},{"cited_title":"Learning to detect a salient object,","cited_arxiv_id":null,"evidence_quote":"Provides the MSRA-B training set used to train ENFNet, so the benchmark comparisons rest on this training data."}],"review_version":1}